Abstract : Many algorithms have been proposed for revealing the community structure in complex networks. Tests under a wide range of realistic conditions must be performed in order to select the most appropriate for a particular application. Artificially generated networks are often used for this purpose. The most realistic generative method to date has been proposed by Lancichinetti, Fortunato and Radicchi (LFR). However, it does not produce networks with some typical features of real-world networks. To overcome this drawback, we investigate two alternative modifications of this algorithm. Experimental results show that in both cases, centralisation and degree correlation values of generated networks are closer to those encountered in real-world networks. The three benchmarks have been used on a wide set of prominent community detection algorithms in order to reveal the limits and the robustness of the algorithms. Results show that the detection of meaningful communities gets harder with more realistic networks, and particularly when the proportion of inter-community links increases.